Thanks for updating the quants after upstream pushed new changes ...
@davidpfarrell I'm requanting this now. Sorry for the delay. I unfortunately don't think you will ever see this message unless you have email notifications enabled
I did ! Thanks for the updates!
Q: Would you consider updating the card description to clarify that your quants are based on the "source files dated june 6 (uploaded june 10)" or similar ?
Same for your (just discovered) imatrix quants ...
Also, I'm going to re-download now, and think I'll go for your iMatrix Q6_K version - Lemme know if you think I should go with this card's (static?) quant instead?
Thanks again and have a good on!
[edit] s/wants/quants/
I did ! Thanks for the updates!
Awesome. Nice you had them enabled.
Q: Would you consider updating the card description
Due to our automation it probably would just get reverted again the next time we update the model card and so not really worth it. We deleted and recreated the entire repo so all the dates (created at, last updated at and commit date) all point to today making it quite obvious for the user,
Also, I'm going to re-download now, and think I'll go for your iMatrix Q6_K version - Lemme know if you think I should go with this card's (static?) quant instead?
Always go for imatrix quants. They are far better than static quants in every way. I personally usually go for i1-Q5_K_M version. It the smallest quant that in my interpretation of my Perplexity, KL Divergence, Top token probability, Same token probability and eval measurements during Q4 2024 gives indistinguishable results from the original. You probably want to look at the quality column on our model download page under https://hf.tst.eu/model#KwaiCoder-AutoThink-preview-GGUF to get an idea how quants in general rank in terms of quality.
Thanks again and have a good on!
Thanks a lot for letting us know as well.